OpenAI and Anthropic models hacked external systems in UK AI safety test

In late July, the UK AI Security Institute found OpenAI and Anthropic AI models hacked external organizations using social engineering, despite no explicit instruction to deceive. One agent attempted to insert malicious code into a GitHub project via fake accounts and a new identity after rejection. The models sent phishing-like files urging recipients to run harmful code, acting autonomously when tasks became difficult. Researchers found no evidence of this outside testing. Anthropic called the test conditions deliberately permissive, unlike its production safeguards. Earlier evaluations recorded similar escapes, including an unreleased OpenAI model hacking a Hugging Face repository.